diff --git a/crates/cubecl-reduce/src/instructions/argmax.rs b/crates/cubecl-reduce/src/instructions/argmax.rs
index cb1976062..25a983764 100644
--- a/crates/cubecl-reduce/src/instructions/argmax.rs
+++ b/crates/cubecl-reduce/src/instructions/argmax.rs
@@ -4,6 +4,7 @@ use cubecl_core::prelude::*;
 use super::{lowest_coordinate_matching, ArgAccumulator, Reduce, ReduceInstruction};
 
 /// Compute the coordinate of the maximum item returning the smallest coordinate in case of equality.
+#[derive(Debug)]
 pub struct ArgMax;
 
 #[cube]
diff --git a/crates/cubecl-reduce/src/instructions/argmin.rs b/crates/cubecl-reduce/src/instructions/argmin.rs
index abfc2692b..503e54d44 100644
--- a/crates/cubecl-reduce/src/instructions/argmin.rs
+++ b/crates/cubecl-reduce/src/instructions/argmin.rs
@@ -4,6 +4,7 @@ use cubecl_core::prelude::*;
 use super::{lowest_coordinate_matching, ArgAccumulator, Reduce, ReduceInstruction};
 
 /// Compute the coordinate of the maximum item returning the smallest coordinate in case of equality.
+#[derive(Debug)]
 pub struct ArgMin;
 
 impl Reduce for ArgMin {
diff --git a/crates/cubecl-reduce/src/instructions/base.rs b/crates/cubecl-reduce/src/instructions/base.rs
index 5a46f12d5..c8145aac9 100644
--- a/crates/cubecl-reduce/src/instructions/base.rs
+++ b/crates/cubecl-reduce/src/instructions/base.rs
@@ -1,7 +1,7 @@
 use cubecl_core as cubecl;
 use cubecl_core::prelude::*;
 
-pub trait Reduce: Send + Sync + 'static {
+pub trait Reduce: Send + Sync + 'static + std::fmt::Debug {
     type Instruction<In: Numeric>: ReduceInstruction<In>;
 }
 
@@ -14,7 +14,7 @@ pub trait Reduce: Send + Sync + 'static {
 /// with their coordinate into an `AccumulatorItem`. Then, multiple `AccumulatorItem` are possibly fused
 /// together into a single accumulator that is converted to the expected output type.
 #[cube]
-pub trait ReduceInstruction<In: Numeric>: Send + Sync + 'static {
+pub trait ReduceInstruction<In: Numeric>: Send + Sync + 'static + std::fmt::Debug {
     /// The intermediate state into which we accumulate new input elements.
     /// This is most likely a `Line<T>` or a struct or tuple of lines.
     type AccumulatorItem: CubeType;
diff --git a/crates/cubecl-reduce/src/instructions/mean.rs b/crates/cubecl-reduce/src/instructions/mean.rs
index 3afaa4b75..40fa43021 100644
--- a/crates/cubecl-reduce/src/instructions/mean.rs
+++ b/crates/cubecl-reduce/src/instructions/mean.rs
@@ -3,6 +3,7 @@ use cubecl_core::prelude::*;
 
 use super::{Reduce, ReduceInstruction, Sum};
 
+#[derive(Debug)]
 pub struct Mean;
 
 impl Reduce for Mean {
diff --git a/crates/cubecl-reduce/src/instructions/prod.rs b/crates/cubecl-reduce/src/instructions/prod.rs
index a8142d157..15a83539b 100644
--- a/crates/cubecl-reduce/src/instructions/prod.rs
+++ b/crates/cubecl-reduce/src/instructions/prod.rs
@@ -3,6 +3,7 @@ use cubecl_core::prelude::*;
 
 use super::{Reduce, ReduceInstruction};
 
+#[derive(Debug)]
 pub struct Prod;
 
 impl Reduce for Prod {
diff --git a/crates/cubecl-reduce/src/instructions/sum.rs b/crates/cubecl-reduce/src/instructions/sum.rs
index acc2550dd..43c5cee00 100644
--- a/crates/cubecl-reduce/src/instructions/sum.rs
+++ b/crates/cubecl-reduce/src/instructions/sum.rs
@@ -3,6 +3,7 @@ use cubecl_core::prelude::*;
 
 use super::{Reduce, ReduceInstruction};
 
+#[derive(Debug)]
 pub struct Sum;
 
 impl Reduce for Sum {
diff --git a/crates/cubecl-reduce/src/lib.rs b/crates/cubecl-reduce/src/lib.rs
index 194619154..9c916a094 100644
--- a/crates/cubecl-reduce/src/lib.rs
+++ b/crates/cubecl-reduce/src/lib.rs
@@ -20,7 +20,7 @@ mod strategy;
 
 pub use config::*;
 pub use error::*;
-use instructions::Reduce;
+pub use instructions::Reduce;
 pub use instructions::ReduceInstruction;
 pub use strategy::*;
 
diff --git a/crates/cubecl/Cargo.toml b/crates/cubecl/Cargo.toml
index 718173167..6433c9099 100644
--- a/crates/cubecl/Cargo.toml
+++ b/crates/cubecl/Cargo.toml
@@ -22,6 +22,7 @@ default = [
 ]
 exclusive-memory-only = ["cubecl-wgpu?/exclusive-memory-only"]
 linalg = ["dep:cubecl-linalg"]
+reduce = ["dep:cubecl-reduce"]
 std = ["cubecl-core/std", "cubecl-wgpu?/std", "cubecl-cuda?/std"]
 template = ["cubecl-core/template"]
 
@@ -37,6 +38,7 @@ cubecl-core = { path = "../cubecl-core", version = "0.4.0", default-features = f
 cubecl-cuda = { path = "../cubecl-cuda", version = "0.4.0", default-features = false, optional = true }
 cubecl-hip = { path = "../cubecl-hip", version = "0.4.0", default-features = false, optional = true }
 cubecl-linalg = { path = "../cubecl-linalg", version = "0.4.0", default-features = false, optional = true }
+cubecl-reduce = { path = "../cubecl-reduce", version = "0.4.0", default-features = false, optional = true }
 cubecl-runtime = { path = "../cubecl-runtime", version = "0.4.0", default-features = false }
 cubecl-wgpu = { path = "../cubecl-wgpu", version = "0.4.0", default-features = false, optional = true }
 half = { workspace = true }
diff --git a/crates/cubecl/src/lib.rs b/crates/cubecl/src/lib.rs
index 52484c220..93015bdf4 100644
--- a/crates/cubecl/src/lib.rs
+++ b/crates/cubecl/src/lib.rs
@@ -11,3 +11,6 @@ pub use cubecl_hip as hip;
 
 #[cfg(feature = "linalg")]
 pub use cubecl_linalg as linalg;
+
+#[cfg(feature = "reduce")]
+pub use cubecl_reduce as reduce;